A testbed for looking at generalisation in sequence learning. The text is split into sentences at each period (.) and each sentence into words. The words are presented through a category encoder, i.e. with non-overlapping input bits.